{"id":"b34c30eb-310a-4c34-81f0-136ef1484d90","arxiv_id":"1908.06686","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Tail sums of Takagi class functions obey LLN, CLT and LIL exactly when the coefficient sequence decays slowly, while geometric decay gives a nonconstant limit in distribution.","lead":"This paper studies how quickly partial sums of generalized Takagi functions converge to the full function. It gives exact conditions for laws of large numbers, a central limit theorem, and a law of the iterated logarithm, and shows that the classical Takagi function instead has a nonstandard limiting behavior.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LIL upper-bound proof applies Lemma 2.7 with s² while the lemma states Σ c_n²; the stronger variance-proxy inequality is unproved and is what the argument needs.","rationale":"The reader's verdict is CONDITIONAL, and this review identifies an additional, more specific obstacle within the same region of the argument: the strong-law part depends on an omitted but likely correct inequality (Lemma 2.3), while the LIL upper bound depends on Lemma 2.7, whose stated form is inconsistent with its subsequent use. The paper's proof of Theorem 1.1(iv) is therefore not complete as written: either Lemma 2.7 must be corrected to the variance-proxy form and proved, or the Borel-Cantelli block argument in Section 2.4 must be replaced. This does not establish that the theorem is false; the result may be true and repairable, which supports keeping the conditional verdict rather than rejecting outright. The Section 2.3 constant slip noted by the reader (the first term in Var(Q²) is omitted from the displayed bound) is real but harmless for convergence, since the corrected bound (4/3)sup c_j²/Σc_n² still tends to zero under (1.6). The reader and this pass agree that the unproved martingale inequalities are the weakest point, but disagree about which inequality is the most load-bearing: the reader chose Lemma 2.3, while this pass finds the Lemma 2.7 statement/application mismatch more serious because it directly affects the LIL conclusion and involves a factor of twelve in the exponent. The proposed concrete test is an independent re-derivation of the maximal inequality and a small exact exponential-moment check for a two-term coefficient sequence, which would settle whether the strong variance-proxy form used in the proof is actually available.","tokens_in":11871,"tokens_out":27536,"duration_ms":284793,"concrete_test":"Re-derive Lemma 2.7 independently from Azuma's Lemma 2 applied to the forward martingale Y_n = M_{b−n}, with differences bounded by |c_{b−n}|/2, and compare the resulting exponential-maximal inequality with the form used in Section 2.4, namely E[exp(λ max|M_N|)] ≤ 8 exp(λ² s²_a/2). If the re-derivation yields only the weaker Σ c_n²/8 or Σ c_n²/2 bound, then the application in Section 2.4 is invalid as written. To test whether the stronger variance-proxy inequality might still be true, compute the exact distribution of (φ*_1, φ*_2) for c_1=c_2=1, c_n=0 otherwise, using the finite Rademacher expansion truncated at R_4 with the tail bounded by 2^{-3}, and evaluate E[exp(λ max_{1≤N≤2}|M_N|)] for λ=1,2,4; compare with 8 exp(λ² s_1²/2).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1.1(iv), first half, rests on the block-maximal estimate in Section 2.4. Lemma 2.7 is stated as E[exp(λ max_{a≤N≤b}|M_N|)] ≤ 8 exp((λ²/2) Σ_{n=a}^∞ c_n²). In the application, λ is chosen as (1+ε)φ(s²_{N_{k+1}})/s²_{N_k} and the displayed bound is written as 8 exp(λ² s²_{N_k}/2 − λu). Since s²_{N_k} = (1/12)Σ_{n=N_k}^∞ c_n², the lemma as stated would give λ² · 6s²_{N_k} in place of λ² s²_{N_k}/2. Substituting the lemma literally, the Markov exponent becomes (1+ε)²φ²(6/s² − 1/s²) = 5(1+ε)²φ²/s² > 0, so the bound is larger than 1 and the Borel-Cantelli argument collapses. To make the proof work, Lemma 2.7 must hold with the variance proxy s²_a in the exponent, i.e. E[exp(λ max|M_N|)] ≤ 8 exp(λ² s²_a/2). The one-line proof 'by a similar argument to Lemma 2 in [3]' does not establish this stronger form: a direct Azuma/Hoeffding bound for the reversed martingale increments d_n=c_nφ*_n, |d_n|≤|c_n|/2, gives an exponent controlled by Σc_n²/8 = (3/2)s²_a, which is twelve times larger in the variance-proxy sense and does not suffice for the log-log summability needed here. The LIL upper bound is therefore not proven as written; the discrepancy between the stated lemma and its application is an internal inconsistency, not merely a missing detail. The omitted proof of Lemma 2.3 is less concerning because standard Hoeffding for reversed martingales yields it up to constants, but Lemma 2.7 is the load-bearing inequality for Theorem 1.1(iv).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Let f(x) = \\sum_{n=1}^\\infty c_n \\phi^{(n)}(x) be a generalized Takagi function with c_n \\in \\ell^1. The paper studies the tail remainder f - f_N as a random variable on [0,1] under Lebesgue measure. Theorem 1.1 characterizes, in terms of tail sums of c_n and c_n^2, when the normalized remainder (f - f_N)/((1/2)\\sum_{n=N}^\\infty c_n) converges to 1 in L^2 or almost surely, and when the centered remainder satisfies a central limit theorem and a law of the iterated logarithm. Conditions (1.4)--(1.7) are given as sufficient conditions, with (1.4) also necessary and sufficient for the L^2 result. Examples with polynomial and exponential coefficients are worked out. Theorem 1.7 treats geometric coefficients c_n = r^n, for which the normalized tail converges in distribution to a nonconstant limit, namely 2(1-r)r^{-1} f_r, while the Ces\\`aro averages of the normalized tail converge to 1.","tokens_in":12377,"tokens_out":26061,"duration_ms":260674,"significance":"The probabilistic framework for Takagi class functions is elegant, and the characterization in Theorem 1.1(i) is an exact, parameter-free result. The paper applies standard martingale limit theorems --- Hall--Heyde, Scott--Huggins, Azuma, Birkhoff --- to a deterministic approximation problem and gives concrete, falsifiable predictions in Examples 1.3 and 1.4. The geometric-coefficient result in Theorem 1.7 is a useful contrast, showing that the usual law of large numbers fails for the Takagi function itself. However, the proof of the LIL upper bound rests on an inconsistent application of Lemma 2.7, and Remark 1.2 is false as stated. These issues are local but load-bearing; once repaired, the paper would be a solid contribution.","major_comments":[{"comment":"The application of Lemma 2.7 in the proof of the upper LIL is inconsistent with the lemma as stated. Lemma 2.7 gives E[exp(λ max_{a≤N≤b}|Σ_{n=N}∞ c_n φ^{(n)}_*|)] ≤ 8 exp((λ²/2)Σ_{n=a}∞ c_n²), whereas the proof uses the bound 8 exp(λ² s_{N_k}²/2 − λu). Since s_N² = (1/12)Σ_{n=N}∞ c_n², the lemma literally gives 8 exp(6λ² s_{N_k}² − λu). With λ = (1+ε)φ(s²_{N_{k+1}})/s²_{N_k} and u = (1+ε)φ(s²_{N_{k+1}}), the exponent becomes 5(1+ε)²φ²/s²_{N_k} > 0, so the bound does not decay and the Borel--Cantelli argument collapses. The needed inequality, with variance proxy s_a² instead of Σ_{n=a}∞ c_n², is not a consequence of the cited argument from Azuma [3]; a direct bounded-martingale-difference argument gives only an exponent controlled by (3/2)λ² s_a², still three times too large. The upper half of Theorem 1.1(iv) therefore needs either a proof of a stronger maximal inequality under condition (1.6) or a different argument.","section":"§2.4, Lemma 2.7"},{"comment":"The claimed corollary in Remark 1.2 is false as stated. For c_n = n^{-α} with α > 1, conditions (1.6) and (1.7) hold, so the second half of Theorem 1.1(iv) gives limsup_N ± M_N/φ(s_N²) = 1 a.s. Hence along a subsequence |M_N| / Σ_{n=N}∞ c_n² = |M_N|/(12 s_N²) ∼ φ(s_N²)/(12 s_N²) → ∞, not 0. This directly contradicts the asserted limit in Remark 1.2. The remark should be removed or replaced with a statement using the correct normalization, such as a bound involving φ(s_N²).","section":"Remark 1.2"},{"comment":"Lemma 2.3 is a load-bearing exponential inequality for the strong law in Theorem 1.1(ii), but it is stated with the note 'whose proof is quite similar to the original one and is omitted.' Because condition (1.5) involves the exact exponential rate, the proof should be supplied or a precise reference given. A standard Azuma/Hoeffding argument does yield a bound of this type, so the issue is likely fixable, but as written the SLLN relies on an unproved statement.","section":"Lemma 2.3"}],"minor_comments":[{"comment":"The asymptotic relations 1/s²_{N_k} ∼ p^k and s²_{N_{k+1}}/s²_{N_k} → 1/p are stated without derivation; they follow from (1.6), (2.7), and the definition of s_N², but a short explanatory sentence would improve readability.","section":"§2.4, equation (2.8)"},{"comment":"The indexing in the definition of the forward martingale (Y_n, G_n) should be made explicit, since n ranges from 0 to b-a and the indices of M_N and T_N decrease as n increases; the current one-sentence description is correct but terse.","section":"§2.4, paragraph before Lemma 2.7"},{"comment":"The statement that condition (1.7) holds only when 0 < β < 1/2 is correct, but the summability threshold could be stated more explicitly by noting that the relevant series behaves like Σ_N N^{-2+2β}.","section":"Example 1.4"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the normalization mismatch in the LIL proof: Lemma 2.7 as stated cannot deliver the estimate used in Section 2.4, and the stronger form needed is not established. The false Remark 1.2 should also be treated as a substantive error, not a typo, since it contradicts the paper's own LIL for the power-law example. The rest of the proof, including the CLT section and Theorem 1.7, appears sound to me."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's main contributions are real: a sharp L2-law for the tail sum of Takagi class functions, and the geometric-coefficient case where the normalized tail converges to the nondegenerate L_r = 2(1−r)/r · f_r. That part is elegant and solid, built on an exact identity and ergodic theory. The CLT via Hall–Heyde is standard but carefully checked, and the L2 iff condition is clean. No fitted parameters, no circularity. The citation pattern is fine — Kono's reverse martingale structure is used as a black box, but that is legitimate and acknowledged.\n\nThe soft spot is serious. The upper-bound half of the LIL (Theorem 1.1(iv), first half, and the Remark 1.2 corollary) rests on Lemma 2.7. That lemma is stated with variance proxy Σ c_n², but the application in Section 2.4 needs it with s²_N = (1/12)Σ c_n². Substituting the lemma as written into the Markov step gives a positive exponent, not the negative one needed for Borel–Cantelli. The one-line proof 'by a similar argument to Azuma' does not establish the stronger form; a direct Azuma/Hoeffding bound on the reversed martingale differences gives a variance proxy controlled by Σ c_n²/8 = (3/2)s²_N, which is still not enough. So the LIL upper bound is not proven as written. This is an internal inconsistency, not a missing detail, and it is load-bearing.\n\nThe omitted proof of Lemma 2.3 is minor by comparison — standard exponential martingale arguments give it up to constants, and the strong law conclusion would survive with a slightly different constant.\n\nWho this is for: probabilists working on Takagi-type functions or reverse martingale limit theorems. The geometric case result is genuinely worth knowing. The paper deserves a serious referee, but the authors need to fix Lemma 2.7 — either prove a maximal inequality with the correct s² proxy, or restructure the LIL argument so it only needs what can be proved. I'd send it to peer review, and I'd expect major revision.","headline":"Nice results on Takagi-class tail sums, including a clean geometric-coefficient phase transition, but the LIL upper-bound proof has a load-bearing variance-proxy mismatch that must be fixed.","tokens_in":12855,"tokens_out":3747,"would_cite":false,"duration_ms":37896,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","60F15","26A27","60G42"],"pacs":[],"model":"deepseek-v4-flash","headline":"For Takagi class functions, the paper derives precise probabilistic rates of convergence: coefficient conditions give L2 and strong laws, a central limit theorem, and a law of the iterated logarithm, while geometric coefficients produce a…","keywords":["Takagi function","Takagi class","rate of convergence","reverse martingale","law of large numbers","central limit theorem","law of the iterated logarithm","Rademacher functions"],"falsifier":"Compute the empirical $L^2$ ratio for $c_n=n^{-\\alpha}$ with $\\alpha>1$: Theorem 1.1(i) predicts $E[((f-f_N)/(\\tfrac12\\sum_{n=N}^\\infty c_n)-1)^2]=\\tfrac13\\sum_{n=N}^\\infty c_n^2/(\\sum_{n=N}^\\infty c_n)^2$, which decays like $((\\alpha-1)^2)/(3(2\\alpha-1))\\,N^{-1}$; a mismatch in the constant or the rate would refute the $L^2$ claim. Separately, for $r=1/2$, the histogram of the normalized tail over uniformly random points should approach the nonconstant distribution of $2T(x)$ rather than concentrating at $1$, testing Theorem 1.7.","tokens_in":11679,"feed_emoji":"📉","tokens_out":16339,"duration_ms":144699,"temperature":0.7,"pith_summary":"The paper asks how quickly the partial sums $f_N(x)=\\sum_{n=1}^{N-1} c_n \\phi^{(n)}(x)$ of a generalized Takagi function converge to $f(x)=\\sum_{n=1}^\\infty c_n \\phi^{(n)}(x)$, treating the tail $f-f_N$ as a random variable on the Lebesgue space $[0,1]$. The main theorem gives coefficient conditions under which the normalized tail $\\bigl(f(x)-f_N(x)\\bigr)/\\bigl(\\tfrac12\\sum_{n=N}^\\infty c_n\\bigr)$ tends to $1$ in the usual probabilistic senses: in $L^2$ if and only if the square tail is small relative to the squared tail sum, almost surely under a stronger exponential condition, and with central-limit and law-of-the-iterated-logarithm fluctuations under further conditions. The paper then shows the opposite behavior for geometric coefficients $c_n=r^n$: the normalized tail converges in distribution to a nonconstant random variable $L_r=2(1-r)r^{-1}f_r$, so the classical Takagi function itself does not satisfy the usual law of large numbers. The upshot is a criterion for when partial-sum approximation is typical at the scale of half the coefficient tail, and when the error keeps random fluctuations of that same order.","feed_headline":"Takagi function's tail fails the law of large numbers","feed_subtitle":"For geometric coefficients the tail error stays random; under other conditions it obeys LLN, CLT, and LIL.","key_machinery":"The central object is the centered tail sum $M_N=\\sum_{n=N}^\\infty c_n\\phi_*^{(n)}$, where $\\phi_*^{(n)}=\\phi^{(n)}-\\tfrac12$ is uniform on $[-1/2,1/2]$. Because the tent map preserves Lebesgue measure, $M_N$ is a reverse martingale—a martingale-like sum adapted to the decreasing tail $\\sigma$-fields generated by the Rademacher functions—and its differences are orthogonal, with variance $\\tfrac1{12}c_n^2$ and fourth moment $\\tfrac1{80}c_n^4$. Variance computations turn the $L^2$ theorem into a ratio of tail sums, while an exponential tail inequality for the reverse martingale (Lemma 2.3) drives the strong law and the upper law of the iterated logarithm. For geometric coefficients the mechanism is different: the exact identity of Lemma 3.1 rewrites the normalized tail as $f_r(2^{N-1}x)/E[f_r]$, and the measure-preserving, ergodic doubling map $x\\mapsto 2x\\bmod 1$ converts this into a fixed nonconstant limiting distribution and, by the ergodic theorem, Cesàro convergence to the constant $1$.","core_discovery":"On the paper's own terms, the central result is Theorem 1.1. Define the centered tail $M_N=f-f_N-\\tfrac12\\sum_{n=N}^\\infty c_n=\\sum_{n=N}^\\infty c_n(\\phi^{(n)}-\\tfrac12)$. The functions $\\phi^{(n)}-\\tfrac12$ are centered and uniformly distributed on $[-1/2,1/2]$, and $M_N$ is a reverse martingale with orthogonal differences $d_n=c_n(\\phi^{(n)}-\\tfrac12)$, so $s_N^2=E[M_N^2]=\\tfrac1{12}\\sum_{n=N}^\\infty c_n^2$. Theorem 1.1 says: (i) $M_N/\\tfrac12\\sum_{n=N}^\\infty c_n\\to 0$ in $L^2$ if and only if condition (1.4) holds; (ii) under (1.5) the same ratio tends to $0$ almost surely; (iii) under (1.6) $M_N/\\sqrt{s_N^2}$ converges in distribution to the standard normal; and (iv) the upper law of the iterated logarithm holds under (1.6), while the full two-sided version with norming $\\varphi(s_N^2)$ holds under (1.7). For geometric coefficients $c_n=r^n$, Theorem 1.7 proves the exact self-similarity identity $\\bigl(f_r-f_{r,N}\\bigr)/\\bigl(\\tfrac12\\sum_{n=N}^\\infty r^n\\bigr)=f_r(2^{N-1}x)/E[f_r]$, which makes the normalized tail have the nonconstant distribution of $f_r/E[f_r]$ for every $N$; ergodicity of the doubling map then gives Cesàro convergence of these normalized tails to $1$ almost surely and in $L^1$.","pith_inferences":["A natural next step is to determine whether the exponential condition (1.5) is necessary for the strong law, and whether Lemma 2.3 can be proved directly for arbitrary $\\ell^1$ coefficient sequences rather than imported from the classical martingale argument.","The same two-case dichotomy—reverse-martingale limit theorems versus self-similar distributional limits—should apply to other continuous functions built from iterated maps, such as Takagi-type sums with slowly varying coefficient blocks.","Because $f_{1/4}(x)=x(1-x)$, the limiting variable $L_{1/4}=6x(1-x)$ has an explicit distribution; exploring other rational $r$ could yield explicit laws for $L_r$ and link the distributional limit to familiar beta-type distributions.","For plotting or numerical approximation of geometric-coefficient Takagi functions, the distributional limit implies an irreducible random-looking vertical error on the scale of the entire tail mean, so error bounds for graphs of $f_N$ should be quantiles of $L_r$, not pointwise intervals around $1$."],"forward_implications":["For coefficients $c_n=n^{-\\alpha}$ with $\\alpha>1$, all hypotheses of Theorem 1.1 are satisfied: the normalized tail obeys both laws of large numbers, and after rescaling by $\\sqrt{N}$ it has a standard normal limit with explicit constant $(\\alpha-1)/\\sqrt{3(2\\alpha-1)}$.","For subexponential coefficients $c_n=e^{-Kn^\\beta}$ with $K>0$ and $0<\\beta<1$, the weak and strong laws and the central limit theorem hold, but the full law of the iterated logarithm is obtained only for $\\beta<1/2$; the paper leaves open whether the gap $1/2\\le\\beta<1$ can be filled.","For geometric coefficients, the normalized tail converges in distribution to a nonconstant random variable, so no deterministic normalization makes it concentrate at $1$; however, the Cesàro average of the normalized tails tends to $1$, giving an ergodic sense in which the half-tail sum is the correct average scale.","The $L^2$ equivalence in Theorem 1.1(i) gives a practical criterion: partial sums approximate the full function at the scale $\\tfrac12\\sum_{n=N}^\\infty c_n$ exactly when the square tail is dominated by the squared tail sum."],"supporting_citations":[{"why":"Supplies the reverse-martingale representation of the centered tail and the multiplicative-system orthogonality used in all variance computations.","marker":"[9]"},{"why":"Provides the classical exponential inequality that the paper adapts, with proof omitted, for the strong law and the upper law of the iterated logarithm.","marker":"[3]"},{"why":"Supplies the reverse-martingale central limit theorem (Corollary 3.4) and the martingale convergence result used to verify the conditions for the law of the iterated logarithm.","marker":"[6]"},{"why":"Quoted as the reverse-martingale law of the iterated logarithm used to obtain the full two-sided version.","marker":"[10]"},{"why":"Gives the measure-preserving and ergodic properties of the doubling map used to pass from the self-similarity identity to distributional and Cesàro convergence.","marker":"[4]"},{"why":"Characterizes exactly when the series defining a generalized Takagi function converges, so the paper works throughout with absolutely summable coefficients.","marker":"[7]"}],"fun_headline_variants":["For Takagi, LLN fails; CLT and LIL hold under conditions","Takagi's tail error never vanishes: no LLN, but CLT and LIL","Takagi breaks LLN; its generalizations obey CLT and LIL","Takagi function itself fails the law of large numbers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is the quoted reverse-martingale representation of the centered tail together with the unproved exponential tail estimate of Lemma 2.3; if either needs extra hypotheses on the coefficients beyond $\\ell^1$, the strong-law and law-of-the-iterated-logarithm conclusions are not established.","fun_headline_variants_meta":{"raw":{"variants":["For Takagi, LLN fails; CLT and LIL hold under conditions","Takagi's tail error never vanishes: no LLN, but CLT and LIL","Takagi breaks LLN; its generalizations obey CLT and LIL","Takagi function itself fails the law of large numbers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000402,"raw_usage":{"total_tokens":2126,"prompt_tokens":1005,"completion_tokens":1121,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":1040}},"tokens_in":621,"tokens_out":1121,"duration_ms":11393,"temperature":1.0,"reasoning_tokens":1040,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:37:04.783580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical $L^2$ ratio for $c_n=n^{-\\alpha}$ with $\\alpha>1$: Theorem 1.1(i) predicts $E[((f-f_N)/(\\tfrac12\\sum_{n=N}^\\infty c_n)-1)^2]=\\tfrac13\\sum_{n=N}^\\infty c_n^2/(\\sum_{n=N}^\\infty c_n)^2$, which decays like $((\\alpha-1)^2)/(3(2\\alpha-1))\\,N^{-1}$; a mismatch in the constant or the rate would refute the $L^2$ claim. Separately, for $r=1/2$, the histogram of the normalized tail over uniformly random points should approach the nonconstant distribution of $2T(x)$ rather than concentrating at $1$, testing Theorem 1.7.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the reverse-martingale representation of the centered tail and the multiplicative-system orthogonality used in all variance computations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the classical exponential inequality that the paper adapts, with proof omitted, for the strong law and the upper law of the iterated logarithm."},{"cited_title":"and Heyde, C","cited_arxiv_id":null,"evidence_quote":"Supplies the reverse-martingale central limit theorem (Corollary 3.4) and the martingale convergence result used to verify the conditions for the law of the iterated logarithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Quoted as the reverse-martingale law of the iterated logarithm used to obtain the full two-sided version."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the measure-preserving and ergodic properties of the doubling map used to pass from the self-similarity identity to distributional and Cesàro convergence."},{"cited_title":"and Yamaguti, M","cited_arxiv_id":null,"evidence_quote":"Characterizes exactly when the series defining a generalized Takagi function converges, so the paper works throughout with absolutely summable coefficients."}],"review_version":1}